Stories
Deep reads, production insights, pattern analysis, and the occasional cross-domain connection.
Six Vectors in Two Weeks
Meta's AI agent went rogue. Hackerbot-Claw exploited Microsoft and DataDog's repos. OpenClaw phishing drained $30M from developer wallets. That's three of six agent security incidents in the last two weeks. The pattern is clear, and it should be front-page news.
The Frontier Club Is Down to Three
Six weeks ago, five companies were building frontier AI models. Now there are three. The exits happened faster than anyone predicted, and they tell us something important about where AI value actually lives.
When AI Reviews the Reviewers
The Linux kernel has the most rigorous code review process in software. An AI agent just found half the bugs it missed. Here's what that means for every production engineering team.
The $1 Trillion Question: NVIDIA's Bet vs. Enterprise Reality
Jensen just promised $1 trillion in AI hardware orders. Enterprise reports show 77% of agent projects dying in pilot. Someone's math is wrong, or the market is about to look very different than both sides expect.
The Agent Security Stack: From Weekend Project to Docker Partnership in Six Weeks
A developer built a security-first agent tool in a weekend. Six weeks later, Docker called. The agent security stack is crystallizing through the same pattern that produced Kubernetes.
BlackRock's AI Bankruptcy Playbook: Follow the Lender, Not the Builder
The CEO of the world's largest asset manager just told you which AI companies will go bankrupt. He also told you who's going to buy the wreckage. Same person.
The Frontier Club Is Shrinking: Meta's $135B Question
The company spending more on AI than most countries' GDP just admitted it might need to rent someone else's model. The frontier club has fewer members than the market thinks.
The Inference Stack Is Converging
Five different teams, working on five different layers of the inference problem, arrived at the same conclusion this week. The cost of running AI models is about to break downward in ways nobody's pricing models account for.
The Synthetic Web Problem: Your AI Agent Can't Tell Truth From a Well-Ranked Lie
Researchers built fake internets to test whether AI agents can find truth when misinformation ranks first. The agents failed. We're deploying them at Pentagon scale anyway.
The Trust Paradox: Anthropic's Principles Created a Vacuum. Google Filled It in 24 Hours.
Anthropic refused to build weapons for the Pentagon. Google launched Pentagon AI agents the next day. Jeff Dean signed a brief supporting Anthropic anyway.
Model Deprecation Is the Real Vendor Lock-In
I shipped a production AI system in January. The model it ran on was deprecated by March. Here's what I learned about building for a world where your foundation disappears every quarter.
AI VC Just Became Two Markets
AI is 90% of all VC. But there are actually two completely separate markets hiding inside that number, and they reward opposite strategies.